[Cyberpradeep]

Summary

This PR implements --bench local as proposed in #5, allowing Harness² to run recursive adaptation sweeps on an arbitrary local directory of instructions and source documents without requiring JobBench, LAB, or WorkBuddy installations.

Implementation Overview

1. LocalAdapter (harness2/benchmarks/local.py):
   • Task Discovery: Discovers tasks in <tasks-dir>/<domain>/<task_id>/task.md.
   • Domain Mapping: Tasks are mapped to their <domain> parent folder, preserving per-domain experience library growth across user tasks.
   • Zero-Dependency Metadata: Parses optional task.toml via Python 3.12 standard library tomllib (outputs, check, timeout).
   • Workspace Isolation: Every rollout receives a clean, isolated workspace with a fresh copy of files/. Rollout edits never bleed into subsequent rollouts.
   • Substrate Launchers: Concrete launchers for OpenCode (run_argv) and Codex (exec).
   • Answer-Free Post-Selection Scoring: If check is configured, it executes strictly post-selection via harness2-judge. Tasks without a check return status="not_configured" and are reported as ungraded rather than judge failures.
   • Baseline Harness Generation: Automatically generates a minimal, valid c1..c8 starter tree or adopts a user-supplied --base-harness.
2. CLI & Preflight Integration:
   • Integrated --bench local, --tasks-dir, and --base-harness across _common.py, run_sweep.py, judge.py, report.py, and check.py.
   • Decoupled harness2/preflight.py from academic benchmark and third-party runner dependencies.
3. Core Seam Purity:
   • Zero changes inside harness2/core/. The implementation proves the benchmark seam in core/spec.py works as intended.

---

Dependency & Coordination Notes

• Resolves: Proposal: --bench local, run Harness² on your own folder of tasks #5
• Coordination with PR core: include task inputs in the reuse identity of a rollout #76 (Task-Input Identity):
  LocalAdapter implements the standard BenchmarkAdapter task-input interface (task_prompt() and task_files()). It performs rollout-level input fingerprint validation today. Once core: include task inputs in the reuse identity of a rollout #76 is merged into main, scheduler-level task-summary resume invalidation will automatically activate for LocalAdapter with zero code duplication.
• Coordination with PR Add opt-in human feedback for selection and reflection #6 (Human Feedback):
  Task IDs follow the <domain>/<task_id> convention, ensuring seamless out-of-the-box compatibility with PR Add opt-in human feedback for selection and reflection #6's --feedback-dir review notes.

---

Validation

• python -m unittest -v tests.test_local_bench: All 7 unit/integration tests passed in 0.47s (OK).
• Tested harness2 --bench local --tasks-dir <path> --dry-run: correctly discovers domains, tasks, and builds isolated namespaces.
• Preflight and CLI option handling tested across run_sweep, judge, report, and check.
• Documentation added in docs/LOCAL_BENCHMARK.md.